akmaru
commented
1 year ago tosa-to-arith も追加すると、 tosa.const もconversionされるようになった。
./bin/mlir-opt -pass-pipeline="builtin.module(func.func(tosa-to-linalg-named))" conv-bn-relu.mlir | ./bin/mlir-opt -pass-pipeline="builtin.module(func.func(tosa-to-linalg))" | ./bin/mlir-opt -pass-pipeline="builtin.module(func.func(tosa-to-arith))"#map = affine_map<(d0, d1, d2, d3) -> (d3, d0, d1, d2)>
#map1 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
#map2 = affine_map<(d0, d1, d2, d3) -> (d3)>
module {
func.func @conv_bn_relu(%arg0: tensor<1x224x224x3xf32>, %arg1: tensor<64x3x3x3xf32>, %arg2: tensor<64xf32>) -> tensor<1x224x224x64xf32> {
%cst = arith.constant 0.000000e+00 : f32
%padded = tensor.pad %arg0 low[0, 1, 1, 0] high[0, 1, 1, 0] {
^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
tensor.yield %cst : f32
} : tensor<1x224x224x3xf32> to tensor<1x226x226x3xf32>
%cst_0 = arith.constant dense<[1, 2, 3, 0]> : tensor<4xi64>
%0 = tensor.empty() : tensor<3x3x3x64xf32>
%1 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg1 : tensor<64x3x3x3xf32>) outs(%0 : tensor<3x3x3x64xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<3x3x3x64xf32>
%2 = tensor.empty() : tensor<1x224x224x64xf32>
%cst_1 = arith.constant 0.000000e+00 : f32
%3 = linalg.fill ins(%cst_1 : f32) outs(%2 : tensor<1x224x224x64xf32>) -> tensor<1x224x224x64xf32>
%4 = tensor.empty() : tensor<1x224x224x64xf32>
%5 = linalg.conv_2d_nhwc_hwcf {dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%padded, %1 : tensor<1x226x226x3xf32>, tensor<3x3x3x64xf32>) outs(%3 : tensor<1x224x224x64xf32>) -> tensor<1x224x224x64xf32>
%6 = linalg.generic {indexing_maps = [#map2, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2, %5 : tensor<64xf32>, tensor<1x224x224x64xf32>) outs(%4 : tensor<1x224x224x64xf32>) {
^bb0(%in: f32, %in_3: f32, %out: f32):
%9 = arith.addf %in, %in_3 : f32
linalg.yield %9 : f32
} -> tensor<1x224x224x64xf32>
%cst_2 = arith.constant dense<0.000000e+00> : tensor<1x224x224x64xf32>
%7 = tensor.empty() : tensor<1x224x224x64xf32>
%8 = linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%6, %cst_2 : tensor<1x224x224x64xf32>, tensor<1x224x224x64xf32>) outs(%7 : tensor<1x224x224x64xf32>) {
^bb0(%in: f32, %in_3: f32, %out: f32):
%9 = arith.maxf %in, %in_3 : f32
linalg.yield %9 : f32
} -> tensor<1x224x224x64xf32>
return %8 : tensor<1x224x224x64xf32>
}
}
TOSA -> Linalg へのconversion
下記の内容を
conv-relu.mlirとして保存した。下記のコマンドで linalg へのconversion
下記のようなIRが得られた。